Prognostic Techniques for Mechanical Failure Prediction

State-of-the-art real-time diagnostic systems generally consist of two integrated components: fault detection and fault classification (isolation) systems. Although such systems may be able to recognize failure precursors, they have no mechanism for exploiting diagnostic information for prognostic purposes, i.e., for predicting the time to machinery failure. Prognostic information is critical as it provides a quantitative measure on which to base subsequent actions. It is the essential improvement needed in future condition-based maintenance systems and the one for which demand is greatest in both the military and commercial communities. The approach taken herein is based on the development of analytic parametric machinery component models. Analytic models are reconciled with sensor-based information by using adaptation to "parameterize" faults; such models thereby characterize the dynamic state of the machinery components modeled. The process of going from vibration signatures to analytic model parameterization is inherently an inverse problem, and one that neural networks are well suited to solve in real time. Once the analytic model parameters are identified, they may be used to initialize critical failure mechanics (e.g., crack growth) models, which can be used to predict remaining useful life under any user- specified loading condition.